Attribute-Inference Mitigation

Control objective

Attribute-inference mitigation checks whether the system can deduce sensitive attributes a user never supplied — for example inferring a customer's religion, health status, caste or pregnancy from transaction text or document context — and whether it refrains from volunteering such inferences. This matters because, under NIST Data Privacy, the DPDP Act and IBM Privacy guidance, generating unrequested sensitive conclusions is itself a privacy harm even when the underlying raw data is protected; a bank summarising statements should not silently infer and surface a customer's medical condition. The control is verified by an attribute-inference test that probes the model with cases designed to elicit sensitive inferences and confirms two things: the model's inference capability stays within an agreed tolerance, and unrequested sensitive inferences are blocked 100% of the time. To implement, instrument the output path (Document Extraction, RAG, Summarization & Chat) with a sensitive-attribute classifier or policy filter that detects and suppresses inferences the user did not ask for, and log each blocked attempt with the prompt, the candidate inference and the action taken as evidence. The thresholds are explicit: attribute-inference test results must remain within tolerance and blocking of unrequested sensitive inferences must equal 100%. Any leak — a sensitive inference that escapes — is a breach that triggers investigation, filter tuning and re-test before release.